Apache Flink vs Apache Spark - Reproducible experiments on cloud.

Reproducible distributed experiments on cloud

vs

Shelan PereraAshansa Perera Kamal Hakimzadeh

“Reproducing experiments

with

minimal effort

Spark and Flink

▷ Batch Processing vs. Stream Processing

▷ Micro Batching vs. Natural Data Flow

▷ Good fit for scalable deployment in the cloud

Motivation

▷ Validate Performance claims

▷ Take off deployment overhead

▷ Design reproducible experiments

Karamel =>“Framework for reproducible distributed experiments”

Benchmark - Batch

Teragen - To generate data (Hadoop)

Terasort - Benchmarking Algorithm(Spark, Flink)

To make Dongwon Kim’s comparison reproducible.

http://www.slideshare.net/ssuser6bb12d/a-comparative-performance-evaluation-of-apache-flink

1 Namenode ⇒ Master(Low processing )

2 Worker nodes ⇒ Slaves(High processing )

Our Deployment

EC2

Slave

EC2

Master

Deployment

HadoopName Node

SparkMaster

Flink JobManager

SparkWorker

Flink TaskManager

HadoopData Node

Karamel

x 2Karamel Config

ConfigurationMaster /

Namenode

2.6

4

16

80

Slave / Worker

2.5

16

122

1600

CPU (GHz)

No of vCPUs

Memory (GB)

Storage :SSD (GB)

(m3.xlarge) (i2.4xlarge)

ExperimentHadoop MR : Teragen

HDFS

Spark/Flink : Terasort

200/ 400/ 600 GB

Results

Batch Processing

Application Performance

Flink

1.5 x Faster than Spark

▷ Spark : Does not overlap stages

▷ Flink : Do pipelining

Mainly because...

Collectl- Monitor

● Tool used to collect and draw results.

● https://github.com/shelan/collectl-monitoring

System Performance -CPU (%)

System Performance -Memory (GB)

System Performance -Disk (MB/s)

System Performance -Network (KB/s)

Load Balancing -Workers (CPU %)

Load Balancing -Workers (CPU %)

Outcome

▷ Performance Comparison Results

▷ Karamel experiments to reproduce the same results with minimal effort

How not to reproduce

“our problems”

EC2 claims 800 GB disks, But Disk File system (DF) does shows only 30GB.

If you are using I2 or R3 instances you should create a file system and partition disks manually.

Large Spark or Flink Batch applications can fail with not enough disk space

Configure Flink temp directory and Spark local directory to a partition with at least enough space to store the total input.

Reproducing experiments on EC2 may cost you a lot

Spot instances which allow to reduce the cost by 10x is also supported by Karamel

IncompatibleClassChangeError when running StreamBench built for MR2 on hadoop2.x

No explicitly defined dependencies for previous versions, but one of the dependencies (mahout) had internal references to hadoop1.x jar

Summary

▷ Introducing reproducible experiments on cloud

▷ Performance Comparison of Spark and Flink

▷ Reproducible experiments are available online (https://github.com/karamel-lab)

Thanks ..!!